5-4 图机器学习在蚂蚁推荐业务中的应用.pdf

1、图学习在蚂蚁推荐中的应用涂珂|自我介绍2011-2015 清华大学计算机系本科2015-2020 清华大学计算机系博士，研究方向：图学习2020-至今：蚂蚁智能引擎技术事业部-金融机器智能-图学习组主要研究方向：图机器学习，通过算法赋能营销ROI、搜索推荐等业务场景。01背景背景02基于知识图谱的推荐基于知识图谱的推荐03基于社交和文本的推荐基于社交和文本的推荐04基于跨域的推荐基于跨域的推荐目录目录CONTENT|背景01|背景基金推荐消费券推荐腰封推荐支付宝上有大量的推荐场景|背景社交网络UU关系推荐网络UI关系图谱网络II关系用户基金板块经理人股票跨域网络UI关系用户基金支付宝域内推荐场

2、景稀疏行为，低活目标基金搜索，无/低理财用户消费券推荐，无/低线下支付用户用户自身的行为数据UI关系很少图谱 II 关系社交UU关系和文本语义IW关系相似场景的推荐UI关系基于图谱的推荐02|Tu K,Cui P,Wang D,et al.Conditional graph attention networks for distilling and refining knowledge graphs in recommendation,CIKM 2021|Knowledge Graph for Recommendation User-Item Graph Sparse New arrival

3、user Side Information attribute contexts images Knowledge graphgenregenredirectedstarredgenrestarredstyleincludeincludestarfriendKnowledge GraphRecommenderSystemPredicted Items!#$%ComedyActioner|Challengesgenregenredirectedstarredgenrestarredstyleincludeincludestarfriend!#$%ComedyActioner Knowledge

4、Graph Distillation most knowledge relationships maybe not helpful=time-consuming and noisy Knowledge Graph Refinement different users may have different preference on different nodes in KG=User1 likes Actors,User2 likes Comedy|Existing methodsEmbedding-based modelslearn entity and relation represent

5、ations first by knowledge graph embedding algorithmsfails to solve the distillation and refinement issuesPath-based methodsdistill the knowledge graphs into multiple meta-pathsheavily depend on the hand-crafted designcan not handle the diversity of the users preference on the knowledge relationships

6、GCN-based methodsuse an attention mechanism to produce weights of different neighbors for knowledge graph refinementonly based on the two nodes in the same edges and are independent with the target nodes|Proposed ModelKGE layerKnowledge GraphDistillation moduleKnowledge GraphRefinement moduleMLP pre